One model, two circuits: own iron and cloud

I wrote it before: the model is a base good, the delivery channel doesn't change quality, the value lives in the layer on top. Now it's two live circuits, each with its own infrastructure partner. One Qwen3.6, two ways to deliver it.

→ Own iron — Selectel. Qwen3.6 on dedicated servers: for those who must keep everything inside their own perimeter — law, medicine, state data. The model on-site, physical isolation. → Cloud — Yandex Cloud. We're partners now; the same Qwen3.6 through their API: for the storefront, trials, and clients without their own hardware. Zero CapEx, minutes to spin up.

Both circuits are Russian, 152-FZ-attested. The client picks the circuit, not the model: where the data sits and how much isolation they'll pay for. The assistant behaves identically — I ran both through my benchmark, the numbers match. One quality passport for both circuits.

The layer on top doesn't change: corpus, retrieval, citation to the clause, honest refusal. Only where it lives changes.